Developing a Model to Detect Harmfulness from Open Chat Names and Descriptions

LINE’s AI Services Lab built a moderation model for open chat names and descriptions to help automate harmful-content review.

The team cleaned noisy historical labels and selected Granite Guardian 3.1 2B, an Apache-licensed decoder model from IBM Research, as the base.

Using prompt-based cross-entropy training with LoRA fine-tuning, the model learned to predict both the penalty code and the reason.

The approach shows how a lightweight generative safety model can expand moderation coverage while reducing manual review at scale.